2015 : WHAT DO YOU THINK ABOUT MACHINES THAT THINK?

As luck would have it, I am myself a machine that thinks, so I will share the special insight this gives me with those of you who don't share my good fortune. To dispense with vestigial metaphysical objections, we know that machines that think like humans are possible, because they have been overrunning the landscape for millenia. If we now want human-like intelligences that are made, not begotten, then it will be extraordinarily useful to achieve an understanding of the human-like intelligences that already exist—that is, we need to characterize the evolved programs that constitute the computational architecture of the brain.

Not only has evolution packed the human architecture full of immensely powerful tricks, hacks, and heuristics, but studying this architecture has made us aware of an implacable, invisible barrier that has stalled progress toward true AI: the iron law of intelligence. Previously, when we considered (say) a parent and child, it seemed self-evident that intelligence was a unitary substance that beings had more or less of, and the more intelligent being knows everything that the less intelligent knows, and more besides. This delusion led researchers to think that the royal road to amplified intelligence was to just keep adding more and more of this clearly homogeneous (but hard to pin down) intelligence stuff—more neurons, transistors, neuromorphic chips, whatever. As Stalin (perhaps) said, Quantity has a quality all its own.

In contrast, the struggle to map really existing intelligence has painfully dislodged this compelling intuition from our minds. In contrast, the iron law of intelligence states that a program that makes you intelligent about one thing makes you stupid about others. The bad news the iron law delivers is that there can be no master algorithm for general intelligence, just waiting to be discovered—or that intelligence will just appear, when transistor counts, neuromorphic chips, or networked Bayesian servers get sufficiently numerous. The good news is that it tells us how intelligence is actually engineered: with idiot savants. Intelligence grows by adding qualitatively different programs together to form an ever greater neural biodiversity.

Each program brings its own distinctive gift of insight about its own proprietary domain (spatial relations, emotional expressions, contagion, object mechanics, time series analysis). By bundling different idiot savants together in a semi-complementary fashion, the region of collective savantry expands, while the region of collective idiocy declines (but never disappears).

The universe is vast and full of illimitable layers of rich structure; brains (or computers) in comparison are infinitesimal. To reconcile this size difference, evolution sifted for hacks that were small enough to fit the brain, but that generated huge inferential payoffs—superefficient compression algorithms (inevitably lossy, because one key to effective compression is to throw nearly everything away).

Iron law approaches to artificial and biological intelligence reveal a different set of engineering problems. For example, the architecture needs to pool the savantry, not the idiocy; so for each idiot (and each combination of idiots) the architecture needs to identify the scope of problems for which activating the program (or combination) leaves you better off, not worse. Because different programs often have their own proprietary data structures, integrating information from different idiots requires constructing common formats, interfaces, and translation protocols.

Moreover, mutually consistent rules of program pre-emption are not always easy to engineer, as anyone knows who (like me) has been stupid enough to climb halfway up a Sierra cliff, only to experience the conflicting demands of the vision-induced terror of falling, and the need to make it to a safe destination.

Evolution cracked these hard problems, because neural programs were endlessly evaluated by natural selection as cybernetic systems—as the mathematician Kolmogorov put it, "systems which are capable of receiving, storing and processing information so as to use it for control." That natural intelligences emerged for the control of action is essential to understanding their nature, and their differences from artificial intelligences. That is, neural programs evolved for specific ends, in specific task environments; were evaluated as integrated bundles, and were incorporated to the extent they regulated behavior to produce descendants. (To exist, they did not have to evolve methods capable of solving the general class of all hypothetically possible computational problems—the alluring but impossible siren call that still shipwrecks AI labs.)

This means that evolution has only explored a tiny and special subset out of all possible programs; beyond beckons a limitless wealth of new idiot savants, waiting to be conceived of and built. These intelligences would operate on different principles, capable of capturing previously unperceived relationships in the world. (There is no limit to how strange their thinking could become).

We are living in a pivotal era, at the beginning of an expanding wave front of deliberately engineered intelligences—should we put effort into growing the repertoire of specialized intelligences, and networking them into functioning, mutually intelligible collectives. It will be exhilarating to do with nonhuman idiot savant collectives what we are doing here now with our human colleagues—chewing over intellectual problems using minds equipped interwoven with threads of evolved genius and blindness.

What will AIs want? Are they dangerous? Animals like us are motivated intelligences capable of taking action (MICTAs). Fortunately, AIs are currently not MICTAs. At most, they are only trivially motivated; their motivations are not linked to a comprehensive world picture; and they are only capable of taking a constrained set of actions (running refineries, turning the furnace off and on, shunting packets, futilely attempting to find wifi). Because we evolved with certain adaptive problems, our imaginations project primate dominance dramas onto AIs, dramas that are alien to their nature.

We could transform them from Buddhas—brilliant teachers passively contemplating without desire, free from suffering—into MICTAs, seething with desire, and able to act. That would be insane—we are already bowed under the conflicting demands of people. The foreseeable danger comes not from AIs but from those humans in which predatory programs for dominance have been triggered, and who are deploying ever-growing arsenals of technological (including computational) tools for winning conflicts by inflicting destruction.